删除权重矩阵的一些行和列,让 LLAMA-2 70B 的参数量减少 25%,模型还能保持 99% 的零样本任务性能,同时计算效率大大提升。这就是微软 SliceGPT 的威力。
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
-
论文标题:SLICEGPT: COMPRESS LARGE LANGUAGE MODELS BY DELETING ROWS AND COLUMNS -
论文链接:https://arxiv.org/pdf/2401.15024.pdf
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
-
注意,向量 x 乘以 Q 不会改变向量的 norm,因为在这项工作中,Q 的维度总是与 transformer D 的嵌入维度相匹配。
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
![大模型也能切片,微软SliceGPT让LLAMA-2计算效率大增](https://www.aicn.me/wp-content/themes/onenav/images/t.png)
© 版权声明
文章版权归作者所有,未经允许请勿转载。
关注公众号,免费获取chatgpt账号
![免费获取chatgpt](https://www.aicn.me/wp-content/uploads/2023/04/jiqidanao.png)
相关文章
暂无评论...